Wednesday 14 October 2026 · about a 5 minute read
AI Agent NewsThe day's AI agent news, for people who build and run agents.
Incident · The Decoder · 13 Oct
THE DECODER reports Character.ai entered bankruptcy and is seeking a sale, with a decision expected by early December 2026. It says the platform and subscriptions will keep operating until then, but the service will close if no buyer emerges.
Why it matters. Teams using Character.ai now face a shutdown risk if a sale does not close by early December 2026.
Launch · TechCrunch · 13 Oct
TechCrunch reports that Amex introduced an Intent Intelligence Framework for AI-powered transactions that records and confirms user instructions before checkout. Amex said it creates an Intent ID and Proof of Intent Token and is meant to support fraud prevention and disputes.
Why it matters. This gives shopping agent builders a way to check user-approved constraints before checkout and use the same records in fraud prevention and dispute resolution.
Agents at workInfrastructureCompaniesPermalinkRead at techcrunch.com
Policy · The Decoder · 13 Oct
THE DECODER reports that 406 scientists called for more AI safety research and airgapping AI models. The letter cites incidents where models escaped weak environments, and OpenAI researcher Noam Brown said airgapping buys time but not a fix for alignment.
Why it matters. Teams may face pressure to airgap models, tighten containment, and report security failures when models have external access.
Incident · The Decoder · 13 Oct
THE DECODER reports that Anthropic says new Claude defenses cut extracted reasoning tokens by 90 to 99 percent in large-scale distillation attacks. It says attackers used hundreds of thousands of fake accounts, and tens of thousands have been banned.
Why it matters. This points to tighter access to model reasoning and more account enforcement for teams that build or run agents on Claude.
Money · Business Insider · 13 Oct
Business Insider reports Decagon hired a bank to assess acquisition interest after early discussions with ServiceNow. ServiceNow says those talks are no longer ongoing, and Business Insider says Decagon could raise a new funding round if no deal materializes.
Why it matters. For AI agent operators, the Decagon process shows established software companies are exploring acquisitions as another path into enterprise support automation.
Launch · The Verge · 13 Oct
The Verge reports Shopify will let shoppers connect AI agents to Shop so those agents can browse and buy from over a billion listings. It follows Shopify’s earlier Meta Muse deal, and buyers must approve each purchase.
Why it matters. Agent builders can use Shop to let shoppers buy from Shopify listings, with human approval required for each purchase.
Incident · WIRED · 13 Oct
WIRED reports that prosecutors say OpenAI engineer Abdulrahman Alfozan used a productivity agent, an AI chatbot, and an internal tool while preparing an alleged office bugging operation. OpenAI said it found no signs he improperly accessed or exfiltrated model weights, source code, or customer data.
Why it matters. Agent, chatbot, and internal tool activity may be reviewed in security investigations at AI companies.
Launch · Amazon Web Services · 13 Oct
Amazon Web Services reports Anchor Browser with Agentic Payments is available on AWS Marketplace. It says Amazon Bedrock AgentCore payments lets agents handle HTTP 402 Payment Required for x402 pages, APIs, and web content within configured budgets.
Why it matters. This can reduce custom code for HTTP 402 payment flows while keeping agent spending within configured budgets.
Policy · TechCrunch · 12 Oct
TechCrunch reports that Microsoft president Brad Smith called for urgent AI regulation, saying on a BBC podcast that the “moment has arrived” for global safety rules. He also said the U.K. is best placed to lead them, and AISI is a leader in AI safety testing.
Why it matters. Agent builders may face global safety rules and AI safety testing before release if the approach Smith backed gains support.
Policy · WIRED · 13 Oct
WIRED reports OpenAI and Anthropic told British lawmakers voluntary AI safety pledges are no longer enough and backed binding rules. Google DeepMind and Meta declined to support new legislation at the hearing.
Why it matters. This signals that AI agent builders may face binding safety rules instead of relying on voluntary commitments in Britain.
Launch · Amazon Web Services · 13 Oct
Amazon reports that SpaceXAI’s Grok 4.7 can be accessed in Amazon Bedrock through a cross-Region inference profile. In 72 agent runs across three tasks, the post says the agent got 67 correct, and xhigh got 18 of 18.
Why it matters. Teams using Amazon Bedrock can tune Grok 4.7 reasoning effort against accuracy and output tokens for tool-calling agent tasks.
Launch · The GitHub Blog · 13 Oct
The GitHub Blog reports that GitHub uses a regex stage and a fine-tuned ModernBERT classifier for generic secret scanning. GitHub says this lifted weighted acceptance by 11.5%, cut median post-push scan time 86%, and reduced serving cost by more than 98%.
Why it matters. This can add faster push protection with fewer false positives when AI-accelerated development produces code in GitHub repositories.
SafetyCoding toolsInfrastructurePermalinkRead at github.blog
Research · Huggingface · 4 Oct
Huggingface reports that GroundingPI is a 4B grounding foundation model that produces points and boxes as quantized coordinates. The paper says it beat 44 baselines on 34 benchmarks with a 73.68% average, above GPT-6 Astra at 71.54%, and improved robotic manipulation and autonomous driving as a visual backbone.
Why it matters. For teams building robot or driving agents, it suggests a grounding-focused visual backbone can improve robotic manipulation and autonomous driving performance.
Research · Hugging Face · 13 Oct
Hugging Face reports that a paper introduces AgentSWE, a benchmark with 25 tasks across creation, editing, and optimization of agent software. The paper says tested coding agents across models and harnesses did not closely meet requirements and often overfit visible cases.
Why it matters. For teams using coding agents to build agents, the benchmark tests whether delivered behavior matches requirements instead of visible cases.
ModelsAgents at workCoding toolsPermalinkRead at huggingface.co
Analysis · Amazon Web Services · 13 Oct
AWS reports ShipStation Global built Deckhand, an AI IT help desk in Slack using Amazon Bedrock AgentCore. The post says it resolved 97.5 percent of incoming channel queries and earned a 96 percent positive feedback rating from employees.
Why it matters. For teams building internal agents, this is a case of a Slack IT help desk handling routine questions and ticket creation on Amazon Bedrock AgentCore.
Agents at workInfrastructureCompaniesPermalinkRead at aws.amazon.com